Carry the reading order across, not just the pages

[tannevaled]

ops v0.6.0 → v0.7.0, reader v0.4.2 → v0.5.0. No code change.

A structure tree is the order a screen reader reads a document in. It is not the
order the marks sit on the page, and nothing else in the file says it.
Plan.Build was writing a document without one, so a tagged form put through a
shared edit came out untagged: still legible to somebody looking at it, and no
longer navigable by somebody who cannot.

Measured

Two binaries, each verified with go version -m. The harness puts every page
of a file into a plan, rotates the first page, adds a bookmark, builds, writes,
reads the result back and reports what the rebuilt catalogue says:

old  ops v0.6.0   reader v0.4.2   pdffont v0.3.0   forms v0.2.1
new  ops v0.7.0   reader v0.5.0   pdffont v0.3.0   forms v0.2.1

Corpus: all 2 268 forms in /Users/Shared/pdfforms. Prevalence is over the
1 633 real documents.

of the 1 633 real documents
source carries a /StructTreeRoot	1 021
rebuilt document carries one, ops v0.6.0	0
rebuilt document carries one, ops v0.7.0	1 014
top-level structure elements carried across	3 095

One catalogue, verbatim, ca-cra/gst190-fill-26e.pdf:

before: catalog-keys=AcroForm,Lang,MarkInfo,Metadata,Outlines,Pages,Type,ViewerPreferences
        structtree=absent
after : catalog-keys=AcroForm,Lang,MarkInfo,Metadata,Outlines,Pages,StructTreeRoot,Type,ViewerPreferences
        structtree=present keys=IDTree,K,ParentTree,ParentTreeNextKey,RoleMap,Type toplevel=1 rolemap=5

It is the whole tree and not an empty shell: the ParentTree, the RoleMap and
the IDTree that make the elements mean anything all come across with it.

Over all 2 268 files: 1 096 build a different document, 1 096 of them larger, 0
smaller, and no file changed whether it opens or whether it builds.

The seven

1 021 sources carry a root and 1 014 come across, so seven do not. All seven have
a root with no /K — a ClassMap, a ParentTree, a RoleMap, and no tree
under them:

uk-govuk/apply-for-help-with-court-and-tribunal-fees__CY_EX160_0625_save.pdf
   source: present keys=ClassMap,ParentTree,ParentTreeNextKey,RoleMap,Type toplevel=0 rolemap=5
   built : absent

There is nothing in them to carry. That is the same shape as an AcroForm
dictionary with no fields in it, which this family already declines to treat as a
form. So 1 014 of 1 014 actual trees are preserved, and I would not want the
other seven "fixed" by writing an empty root into the output.

Not measured

Wall clock and peak memory. Three other corpus jobs were running on this machine,
so a timing figure would be fiction. The output diff is unaffected by load.

Gates

go vet, gofmt, exact 100% statement coverage, nine cross-compile targets, all
pass locally with GOWORK=off CGO_ENABLED=0.